Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

CUDA: no FP16 arithmetic for vector FA kernel (#17558)

Browse Source
This commit is contained in:
Johannes Gäßler
2025-11-28 10:29:09 +01:00
committed by GitHub
parent 35cf8887e1
commit 73955f7d2a
3 changed files with 24 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cuda/common.cuh
+6 -2
View File
@@ -558,8 +558,12 @@ static __device__ __forceinline__ void ggml_cuda_mad(float & acc, const float2 v
acc += v.y*u.y;
}
static __device__ __forceinline__ void ggml_cuda_mad(float & acc, const half2 v, const half2 u) {
#if defined(GGML_USE_HIP) && (defined(RDNA2) || defined(RDNA3) || defined(RDNA4) || defined(__gfx906__) || defined(CDNA))
#define V_DOT2_F32_F16_AVAILABLE
#endif // defined(GGML_USE_HIP) && (defined(RDNA2) || defined(RDNA3) || defined(RDNA4) || defined(__gfx906__) || defined(CDNA))
static __device__ __forceinline__ void ggml_cuda_mad(float & acc, const half2 v, const half2 u) {
#ifdef V_DOT2_F32_F16_AVAILABLE
asm volatile("v_dot2_f32_f16 %0, %1, %2, %0" : "+v"(acc) : "v"(v), "v"(u));
#else
#ifdef FAST_FP16_AVAILABLE
@@ -571,7 +575,7 @@ static __device__ __forceinline__ void ggml_cuda_mad(float & acc, const half2 v,
acc += tmpv.x * tmpu.x;
acc += tmpv.y * tmpu.y;
#endif // FAST_FP16_AVAILABLE
#endif // defined(GGML_USE_HIP) && (defined(RDNA2) || defined(RDNA3) || defined(RDNA4) || defined(GCN5) || defined(CDNA))
#endif // V_DOT2_F32_F16_AVAILABLE
}
static __device__ __forceinline__ void ggml_cuda_mad(half2 & acc, const half2 v, const half2 u) {